Looked but didn't see: inattentional blindness and yes-bias confabulation in vision-language models

The study examined whether vision-language models (VLMs) exhibit inattentional blindness similar to humans when objects are present in their visual field. Researchers tested models with tasks to identify a gorilla in images and videos of lung CT scans. They found that VLMs are capable of spotting the gorilla, but display inattentional blindness that varies by model type and stimulus type. The Gemini-3.1-Pro model outperformed most other models. In segmentation experiments, the specialized medical model BiomedParse incorrectly flagged a gorilla on 82% of frames in control videos without a gorilla. The research indicates that claims about whether a model detected a specific object require signal-detection analysis with matched control baselines to exclude false positives.